Ensemble siamese networks for object tracking

• Published in: Neural computing & applications Vol. 34; no. 10; pp. 8173 - 8191
• Main Authors: Huang, Hanlin, Liu, Guixi, Zhang, Yi, Xiong, Ruke, Zhang, Shaoxuan
• Format: Journal Article
• Language: English
• Published: London Springer London 01.05.2022; Springer Nature B.V
• Subjects: Artificial Intelligence; Artificial neural networks; Clustering; Clusters; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Control theory; Data Mining and Knowledge Discovery; Datasets; Discrimination; Experiments; Image Processing and Computer Vision; Optical tracking; Original Article; Probability and Statistics in Computer Science; Semantics; Training; Visual fields; Siamese network; K-means; Visual tracking; Ensemble learning
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-022-06911-4

In recent years, considering a balanced accuracy and efficiency, Fully-Convolutional Siamese network (SiamFC) is widely used in the field of visual tracking. Although SiamFC has achieved great success, it is still frustrated in discrimination especially in the discriminative scene. The main reason for the poor discrimination ability of SiamFC is that during the training process, it pays more attention to fitting the whole dataset than learning discrimination ability to similar objects. In terms of this issue, we propose Ensemble Siamese networks (ESiamFC) for tracking by introducing ensemble learning into SiamFC. In detail, firstly, we map the training dataset ILSVRC2015 into embedded space. Secondly, we use balanced k-means to cluster video features. Thirdly, in each cluster, we apply transfer learning into SiamFC to obtain k base trackers with their preferences. Last but not least, to leverage the diversity of base trackers, we propose a Cluster Weight fusion module which can automatically assign fusion weight to base trackers according to the semantic information of the tracking object. Extensive experiments on multiple benchmarks demonstrate that our tracker outperforms SiamFC in precision with a relative increase of 7.1%, 8.6%, 6.7% on Tcolor128, DTB70, LaSOT, respectively.